1 / 19

Optimistic Design

Optimistic Design. Guarded Methods. Do something based on the fact that one or more objects have particular states Make a set of purchases assuming all items are available and not too expensive Make a withdrawal based on the fact that the client has sufficient balance How do we do that?.

jeannieh
Download Presentation

Optimistic Design

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Optimistic Design

  2. Guarded Methods • Do something based on the fact that one or more objects have particular states • Make a set of purchases assuming all items are available and not too expensive • Make a withdrawal based on the fact that the client has sufficient balance • How do we do that?

  3. Pessimistic Design • Lock the relevant object(s) to avoid concurrent changes • Check that state is still OK • If the state is no longer as required, release the lock • What do we do next in this case? • Apply the necessary actions • Finally release the lock

  4. Analysis • Consider a scenario in which we lock every access to an object. Assume: • The program has 1,000 objects like ours • One of the other threads is accessing one of those objects 10% of the time • The changes made to the object invalidate our action 10% of the time • On average the synchronization is necessary only 0.001% of the time! • But we are paying for it 100% of the time for each object! • Maybe we can skip the synchronization? • Whatever can go wrong, will go wrong

  5. Optimistic Design • Attempt actions with / without partial synchronization • Rollback (undo) actions that should not have been performed after we discover the problem • Main issues: • How to discover the problem • How to rollback • What to do after the rollback (on failure)

  6. Commit / Verify Commit / Verify Advantages of Optimistic Design • Reduces the time holding locks • Reduces the number of lock operations

  7. When to Apply • Chance of failure is low enough • Synchronization overhead is high enough • Can tell when some operations has been applied incorrectly / illegally • Always can go back to the last correct state

  8. Example In a word processor we have two threads: the user thread and the spell checker thread. The user thread should never wait for the speller thread while it is spell checking a text. The speller thread should take pieces of text and highlight spelling mistakes. Solution: the document is composed of a sequence of text chunks. In order to edit text, the user thread locks the chunk containing the current editing position and marks the chunk as "busy", then it unlocks the chunk. When editing ends, the user thread locks the chunk again, sets a time-stamp that marks the completion of the editing, and marks the chunk as "not busy". Then, it unlocks the chunk. The spell checker looks for threads that are not busy. When it finds one, it locks the chunk, verify that it is indeed "not busy" and if so, makes a replica of the text contained in the chunk. The thread then finds the spelling mistakes and keeps a local records of them. Then, it locks the chunk again. If the chunk is busy or has been updated after the first lock, the spell checker discards its results. Otherwise, it highlights the erronous words and unlocks the chunk.

  9. Provisional Action • Only “pretending” to act, delaying commitment of effects until the possibility of failure has been ruled out • Easier to manage for methods that only update instance variables • Instead of making changes to an object, act on its shadow copy • Once in a while commit our changes by replacing relevant objects with our copies • Most of the time need to make sure the originals had not changed at all! • Using immutable fields makes it easy to recognize changes by comparing references rather than data.

  10. Provisional Action - Code class Optimistic { // State class is used instead of a few fields. // State should be immutable for this to be efficient (and safer) private State currentState; synchronizedboolean commit(State assumed, State next) {// assumed is state before the action boolean success = (currentState == assumed); if (success) // no interference – safe to commit currentState = next; // the new state after action return success; } } Can a thread change the state during the commit of another?

  11. Rollback / Recovery • Undoing the effects of each performed action • Every action performed within an optimistic method should have an inverse action • Keeping a log of all actions performed so far to allow rolling back (undoing) in reverse order • Periodically, verify that the actions were applied correctly (e.g. without interference) and discard parts of the log • Allows concurrent modifications as long as we can check that they were applied correctly

  12. Rollback Code class Optimistic { private State currentState; // State class is now mutable boolean synchronizedverifyExecution(UndoList log) { booleansuccess = checkGuards(currentState); // check for indication that the operations applied correctly if (!success) // apply inverse of each operation in reverse order log.undo(currentState); else log.clear(); // operations effectively committed return success; } }

  13. Shadowing vs Rollback • Acting on a copy is usually MUCH easier to implement • Shadowing does not allow concurrent modifications, even if independent • Can be solved by improving state granularity or providing state merge operation • In rollback there is no guarantee of progress • Without necessary precautions all threads might keep rolling back each other forever • In shadowing at least one thread makes progress

  14. Helper Methods • How to ensure safety when we perform one action from within another, or even recursion? • Shadowing: • Operate on the state copy passed as a parameter: State func(State before) • Main method commit the copy when it is ready • Rollback: • Helper method returns the list of undo actions for the modifications it performed • Caller adds those to its own list and verifies correctness of the object state periodically • Roll back when we discover a problem

  15. Detecting Problems • In Shadowing we only need to verify that the state of the target object has not changed since we started applying our modifications • Provide an efficient comparison operator (simple reference comparison on immutable objects) • If there is too much to compare, use versioning or timestamping • In Rollback we must have a set of well defined verifiable conditions on the object’s state, so that • those conditions are true if and only if all actions performed on the objects so far were performed correctly and consistently

  16. Volatile Members • Marks for the compiler, that the field may be accessed without synchronization • Unlike normal fields, reading / writing goes directly to memory • True current values vs performance penalty • C++’s equivalent to Java’s volatile is atomic<T> - more on that next week.

  17. Failure Handling • Report failure or throw exception • Allows the caller to decide • Internally retry the action until it succeeds • Are we sure it will eventually succeed? • Retry a number of times, or until a timeout • Drop the action because it is no longer relevant • For example avoid retrying to update a user entry if the user has just been deleted

  18. Retrying Smartly • What may cause an optimistic method to fail? • Retrying an optimistic method immediately is not efficient • It is unlikely that object state has changed to a desirable or consistent state • It is unlikely that thread contention has passed • When using shadowing, since we expect failures to be rare, it may be wiser to subscribe for necessary notifications and wait • May be done very efficiently • Additional programming overhead • In either case, it is beneficial to yield and allow the other threads to execute

  19. Exponential Backoff • When using rollback, we may enter livelock, by causing several threads to rollback/retry continuously • Most operations, when executed without interference for sufficiently long time, will complete successfully • Known as obstruction-freedom • Use exponential backoff • After the first failure sleep for a bounded but random number of time units • After each consecutive failure, go to sleep for 1,2,4,8,16… time units

More Related